Vladimir,
Good suggestions! Could you create a JIRA ticket and put your thoughts
there? It makes sense to optimized the portable marshaller in Java as well.
--
Denis
On 9/23/2015 1:38 PM, Vladimir Ozerov wrote:
> Igniters,
>
> I was a bit concerned about portable marshalling performance in .Net and
> performed a very quick and dirty evaluation of the following optimizations:
>
> 1) For non-"marshal-aware" objects metadata is constant and can be
> calculated in advance and only once.
> 2) The same goes for field IDs - we can calculate them only once.
> 3) Some small writes can be merged into a bigger one. E.g. 1 byte header +
> 1 byte user flag + 4 bytes type ID can be written as 1 pre-calculated long.
> In some cases writeInt(fieldId) + writeInt(fieldLen) ==
> writeLong([constant]), etc..
> 4) No need for raw-mode checks for non-"marshal-aware" objects.
> 5) Memory bounds checks can be removed in lots cases when fields are of
> fixed size.
> 6) In extreme scenario of object having all fields of a fixed size it can
> be written as follows: [request enough memory] + [constant_part1] +
> [field1] + [constant_part2] + [field2] + [...].
>
> Deeper work on this can give even more ideas.
>
> When I applied some of these things in .Net it gave me about ~40%
> marshalling throughput increase in particular scenario. For .Net this is
> very important because marshalling is one of the main reasons why it is
> slower than Java.
>
> Being a micro-opt separately, together these things can significantly
> reduce marshalling CPU consumption.
>
> Probably we can use some of these techniques in Java as well.
>
> Vladimir.
>
Locked
